多模态影像融合在神经系统疾病诊断中的 应用进展——基于人工智能的技术范式与临床转化

Advances in the Application of Multi-Modal Image Fusion in the Diagnosis of Neurological Diseases—AI-Based Technical Paradigms and Clinical Translation

摘要: 神经系统疾病具有病理机制复杂、临床异质性显著、早期隐匿性强等特点，传统单一模态影像学检查难以全面反映疾病的整体病理生理过程。多模态影像融合技术通过整合结构、功能、分子等多维度成像信息，突破单一模态的局限性，实现从“形态学描述”向“系统生物学表征”的跨越式发展。近年来，随着大规模多模态神经影像数据库的建立和人工智能算法的快速发展，基于深度学习的多模态融合策略在阿尔茨海默病、帕金森病、脑卒中、自闭症谱系障碍等疾病的早期诊断、进展预测和个体化治疗决策中展现出显著优势。融合范式从传统的早期、中期、晚期融合向基于正交补偿、跨层引导、注意力机制等高级交互模式演进；可解释人工智能技术的引入为破解“黑箱”困境、提升临床信任度提供了有效路径。本文系统综述多模态影像融合技术的最新进展，阐述MRI-PET、结构–功能成像、影像–电生理等多模态组合在神经系统疾病评估中的应用现状，分析数据异质性、模态缺失、模型泛化能力等核心挑战，并展望多模态人工智能在神经疾病精准医疗中的发展前景。

Abstract: Neurological diseases are characterized by complex pathological mechanisms, significant clinical heterogeneity, and strong early concealment. Traditional single-modal imaging examinations are difficult to fully reflect the overall pathophysiological process of the diseases. Multi-modal image fusion technology breaks through the limitations of single modality by integrating multi-dimensional imaging information such as structure, function and molecule, and realizes a leaping development from “morphological description” to “systems biology characterization”. In recent years, with the establishment of large-scale multi-modal neuroimaging databases and the rapid development of artificial intelligence algorithms, deep learning-based multi-modal fusion strategies have shown significant advantages in the early diagnosis, progression prediction and individualized treatment decision-making of Alzheimer’s disease, Parkinson’s disease, stroke, autism spectrum disorder and other diseases. The fusion paradigm has evolved from the traditional early, middle and late fusion to advanced interaction modes based on orthogonal compensation, cross-layer guidance and attention mechanism; the introduction of explainable artificial intelligence technology provides an effective path to break the “black box” dilemma and improve clinical trust. This paper systematically reviews the latest progress of multi-modal image fusion technology, expounds the application status of multi-modal combinations such as MRI-PET, structural-functional imaging and imaging-electrophysiology in the evaluation of neurological diseases, analyzes the core challenges such as data heterogeneity, modality missing and model generalization ability, and prospects the development prospect of multi-modal artificial intelligence in precision medicine of neurological diseases.